Experiment Design for Stochastic Three-Dimensional

Reconfiguration of Modular Robots

Victor Zykov and Hod Lipson
Computational Synthesis Laboratory
Cornell University, Ithaca, NY 14853
Email: vz25@cornell.edu

Two important trends of modern modular

robotics research are module size reduction and
the increase of the total number of modules in a
robotic system. Smaller and simpler modules are
expected to be cheaper if fabricated in bulk and
easier to selectively replace in a modular system in
case of failures. Stochastic methods of modular
robots assembly and reconfiguration enable
individual
module
simplification
and
miniaturization because modules can be
constructed without on-board power supplies,
locomotion controllers, or actuators [1, 2].
However, stochastic reconfiguration is in many
respects dependent upon the properties and
operation of the environment where the modules
reconfigure [3-5]. This paper discusses the
challenges in design and control of the artificial
environment constructed for experiments in threedimensional stochastic reconfiguration of modular
robots.
Introduction
While planar controlled reconfiguration of
stochastic modular robots has been successfully
demonstrated by several studies [4, 1, 3], volumetric
stochastic reconfiguration of multiple robotic
modules is yet to be achieved. There are several
technical challenges along the way to achieving this
goal. First, whereas in planar experiments modules
require sufficient freedom of motion in a plane, threedimensional reconfiguration necessitates providing
modules with a physical ability to freely move in 3D.
There are only few conceivable options of realizing
this without the use of tethers and manipulators: By
carrying out experiments in either microgravity or
neutral buoyancy. Second, three-dimensional
modules require more sophistication in connector
design, as they do not have any common alignment
plane, which all 2D system naturally have. The
support plane acts as an external alignment
mechanism for all 2D stochastic systems, effectively
reducing their connector dimensionality by one

degree of freedom compared to the 3D systems.

Finally, actuation of planar stochastic system is
facilitated by the availability of the third dimension:
for example, in case of reconfiguration on an air
table, the air moves orthogonally to the surface of the
table, and its flow determines the mobility of the
modules. In a three dimensional system, this type of
mobility control is not readily available.
This paper describes one approach to experiment
design in controlled three-dimensional stochastic
reconfiguration of modular robots that addresses the
above stated challenges. The following sections
present the structure of the experimental installation,
the mechanism of assembly and reconfiguration, as
well as the modular robots control organization and
the user interface.
Three-dimensional module flotation through
neutral buoyancy and global system actuation
In order to provide the robotic modules with
sufficient freedom of motion in three dimensions, we
have constructed a neutral buoyancy environment as
shown in Fig. 1. Neutral buoyancy is achieved by
placing the modular robots inside of a tank filled with
mineral oil (Fig. 1a) after adjusting the density of the
modular robot so that it nearly equals the density of
the supporting fluid. Similar to all other stochastic
systems, we use a single global actuator instead of
many individual actuators for every robot. In our
case, this is a 1hp fluid pump (Fig. 1b) that drives the
oil and, consequently, the suspended modular robots
inside of the reconfiguration chamber. The fluid flow
recirculation system (Fig. 1c) is also equipped with
two 3-way ball valves (Fig. 1d) for fluid flow control
and reversing. These valves are necessary for
additional fluid agitation, as well as for the robotic
structure disassembly, as will be described below.
The ball valves are driven with electrical actuators
(Fig. 1e), controlled from by user interface software
via a USB mechanical relay switch (Fig. 1f). Fig. 2
shows the fluid flow recirculation paths for the
forward and reversed flow directions.

Figure 1. Fluidic experimental environment: a) fluid-filled neutral buoyancy tank, b) fluid pump, c) piping for fluid flow circulation, d)
3-way ball valves for reversive fluid flow control, e) electric ball valve actuators, f) USB-controlled valve actuator switch, g) fluid filter,
h) fluid leak control platform, i) robot electrical power supply and communication interface, j) webcam, k) robot buoyancy control
station.

Figure 2. Fluid flow schematics: (left) forward flow, (right)

reversed flow

Even though, at first glance, it appears that the

problem of free motion in 3D has been solved for the
stochastic modules by using a neutral buoyancy
environment, there still remains a serious challenge
of adjusting the weight of the module to make it
exactly neutrally buoyant. After a few minutes of
operation, the temperature of oil starts to increase due
to fluid recirculation. This temperature change affects
the robot buoyancy, presumably due to unbalanced
thermal expansion of the oil and the parts of the
modular robot.
Experimental tank and reconfiguration
mechanism
The fluid tank is shown in Fig. 3. It consists of
three sections: the main chamber (Fig. 3a) where the
robotic modules (Fig. 3b) are placed for experiments

in assembly and reconfiguration, the fluid collector

(Fig. 3g) separated from the main chamber by the
array of 55 modular robot connection ports (Fig.
3c), and the fluid inlet area (Fig. 3d) separated from
the main chamber by the diffuser plate (Fig. 3f). The
array of robot connection ports serves as a substrate
for the growth of modular structures. It provides
the attached modules with both the electrical power
supply and the instructions for further structure
modification
through
the
electrical
and
communications interface (Fig. 3i).
The fluid flow is normally directed from the top
of the tank to its bottom. By selectively opening the
fluid valves of the substrate sockets, the flow can be
redirected towards the locations on the substrate
plate, where the placement of a modular robot is
desired. As long as the neutrally buoyant modules
move along with the fluid in which they are
suspended, the flow of the carrying fluid will drive
them to the selected substrate sockets. This method of
actuation gives some degree of control over the speed
of the reconfiguration process, similarly to the use of
an air table for the 2D stochastic robots. In fact, we
expect that using a pump to drive the fluid towards
the selected substrate locations should accelerate the
assembly process, compared to the case when such
directional driving force is absent.

Figure 5. Symmetric and redundant electrical signal

designation to the terminal pins.

Figure 3. Experimental chamber: a) robot assembly and

reconfiguration area, b) stochastic modules, c) substrate
plate, d) fluid inlet, e) fluid outlet, f) diffuser, g) collector area,
h) safety clamps, i) robot electrical power supply and
communication interface.

After the latches are mechanically engaged,

attached robot can be
disconnected from the
structure by actuating
the shape memory
alloy actuators of the
latches (SMA, see
Fig. 6) and reversing
the
fluid
flow
direction to push Figure 6. Mechanical latch
away the detached with an SMA actuator.
cube.
Reconfiguration process control

As shown in Fig. 3 and Fig. 4, the robotic

substrate sockets and the modular robots are
equipped with geometrically compliant interfaces and
mechanical latches. The surfaces of the robot shells
and the substrate sockets are designed with 4-way
symmetry enabling the freely floating robots to align
themselves passively under the influence of driving
fluid flow onto the substrate or another robots socket
along both unrestricted degrees of freedom.
Once the robots are aligned and pressed towards
each other by the pressure of the fluid, their latches
(see Fig. 4) engage and they become mechanically
and electrically connected. The electrical and
communications interface of the robots is also a 4way symmetrical hermaphroditic spring loaded pinto-pin contact with the pin travel smaller than half of
the pin tip radius. The interface signal designation is
shown in Fig. 5.

We also developed visual user interface software

that provides intuitive means of control over the
process of stochastic object construction. Fig. 7
shows several screenshots of this software. The user
can change the number and location of the robotic
modules in the resulting structure, selectively control
valve and latch actuators for any of the substrate
sockets or the attached modules, adjust or reverse the
fluid flow. The software converts the visual user
commands into the sequences of serial messages that
are sent from the PC into the appropriate robot or
substrate socket controllers.

Figure 4. Stochastic modular robot with and without the

compliance shells.

Figure 7. Visual interface for stochastic reconfiguration

control.

Conclusions

Figure 8. An external ATMega8 microprocessor is used to

convert RS-232 signal from PC into 1-wire signal for the
stochastic modules and vice versa.

For example, a visual interface command of

adding a cube in a given location will result in
opening the fluid valves of the assembled cubes and
substrate sockets adjacent to this location to direct the
fluid flow and the free floating cubes towards the
point of construction. Conversely, a visual command
of removing an assembled cube will result in
disengaging the latches of the cube that is to be
detached, opening the fluid valves of the adjacent
cubes and substrate sockets and reversing the flow to
push this detached robot away from the remaining
structure.
The communication between the PC and the
robot controllers is organized using modified Dallas
1-Wire protocol [6] according to the diagram shown
in Fig. 8. The original protocol provided fully master
controlled data reading and writing routines. Custom
modifications were made to permit multiple masters
to exist on one bus, and, in this way, to enable
physical event-driven logic for the modular robotics
reconfiguration. To provide a standard PC with
access to 1-wire bus, an external ATMega8 chip was
used as a bridge: it was connected to both the robot
reconfiguration chamber with 1-wire bus, and to the
PC via a TTL-adapted RS-232 interface.
An external USB camera is used to observe and
record the processes of modular robot assembly and
reconfiguration. A sample sequence of captured
images of robot reconfiguration is shown in Fig. 9.

Figure 9. Example frames captured by the webcam show

robot alignment and attachment to a substrate socket.

We present an artificial environment constructed

for experimental research in three-dimensional
assembly and reconfiguration of modular robots and
address three important technical challenges
associated with achieving successful reconfiguration:
First, the robots must be provided with full
freedom of volumetric motion and orientation.
Proposed system supports the modular robots in 3D
by making the modules neutrally buoyant in mineral
oil. Even though this statement is quite simple, the
practical difficulty of this particular task is quite high,
and this factor can be prohibitive in achieving
successful assembly and reconfiguration.
The second challenge is that of module
alignment in a higher dimensional space, than for 2D
stochastic systems. Our proposal with this regards is
to first use the force of the directed fluid flow to drive
the floating inactive modules towards the connection
sites, and, when the modules are within a distance of
physical contact from one another, choose their
surface geometry so that it could align them correctly
under the influence of the driving fluid pressure. This
approach also serves as a solution to the third
identified technical problem of global module
actuation.
Acknowledgement
This research was supported in part by the U.S
National Science Foundation grant CMMI 0634652.
References:
[1] Bishop, J., Burden, S., Klavins, E., Kreisberg, R.,
Malone, W., Napp, N., Nguyen, T. , "Self-organizing
programmable parts," in Proc. Int. Conf. Intelligent
Robots Systems, 2005, pp. 36843691.
[2] Zykov, V., Lipson, H., "Fluidic Stochastic Modular
Robotics: Revisiting the System Design," in
Proceedings of Robotics Science and Systems
Workshop on Self-Reconfigurable Modular Robots
Philadelphia, PA, 2006.
[3] Griffith, S., Goldwater, D., Jacobson, J. M. , "Selfreplication from random parts," Nature, vol. 437, p.
636, 2005.
[4] White, P., Kopanski, K., Lipson, H., "Stochastic SelfReconfigurable Cellular Robotics," IEEE
International Conference on Robotics and
Automation, pp. 2888-2893, 2004.
[5] White, P., Zykov, V., Bongard, J., Lipson, H. , "Three
Dimensional Stochastic Reconfiguration of Modular
Robots," in Proceedings of Robotics Science and
Systems MIT, Cambridge, MA, 2005, pp. 161-168.
[6] Atmel Corporation, AVR318: Dallas 1-Wire
master, Rev. 2579A-AVR-09/04 available at
http://www.atmel.fi/dyn/resources/prod_documents/do
c2579.pdf.